Build compiles uniflow-web in CI with submodules; deploy only runs
actions/deploy-pages after the Pages artifact is uploaded.

Document M4 timings (~1.4× realtime on Small Q8) so the README highlights
how fast local inference feels with the default quant packs.

Emscripten 4.0.10's built-in Dawn port lacks InstanceFeatureName /
QueueWorkDone(status, StringView); fetch a compatible package and pass
EMDAWNWEBGPU_DIR (also bump CI emsdk to 4.0.23).

Stock browsers lack WebAssembly.promising, so JSPI WaitAny blew up on
initBackend. Switch ggml/web build to ASYNCIFY and await embind calls.

Show overall/% progress while fetching weights, default to Small Q4_0 for
faster loads, and coerce pointer/token BigInts so Generate does not throw.

Copy samples into a JS Float32Array in getPcm() instead of exposing heap
pointers that break once addresses exceed signed int32 range.

ggml classifies integrated/unified-memory CUDA GPUs as IGPU rather than
GPU, so has_gpu was false on GB10 despite CUDA0 doing the actual DiT/VAE
compute. This caused a misleading "no GPU backend, running on CPU"
warning, skipped GPU tensor-pinning optimizations, and made explicit
GGML_BACKEND=CUDA/GPU overrides silently fall through to auto-select.
Metal/WebGPU always report GPU directly, so this is a no-op on those
backends.

Measured small/base/large Q8_0 on NVIDIA GB10 with the same methodology
as the existing Apple M4 table. Also documents --batch/--output-dir/
--instruction-idx (already supported, missing from the flag table),
corrects the backend note to mention CUDA, links the browser WebGPU
demo from the top-level README, and adds the upstream UniFlow-Audio
BibTeX citation.